Using correspondence analysis to select training set for multi-modal information data

Abstract

In order to integrate talking and reading with an NAO robot to excite the communication willingness of autistic children, the robot should make appropriate book choice. In this paper, a picture book was divided into the textual information and the image information, we proposed a new approach for picture books training set selection, which combined the multi-modal information of each picture book to select representative and integral training samples for capturing sufficient picture books by correspondence analysis. In our provided method, an improved chi-square statistic to get relative terms and an near-duplicated keyframe mining method to get image information are proposed. Finally, the experimental results demonstrated the feasibility. Moreover, our approach can cut down the computational costs and perform well compared with the manual selection method and the method based on the Q-factor analysis.